Seeing through Denoising: Frequency-guided One-step Denoising for Infrared Small Target Detection

Xuelin Qian; Zixuan Wang; Chang Xu; Dingwen Zhang; Junwei Han

cs.CVOct 21, 2026

Seeing through Denoising: Frequency-guided One-step Denoising for Infrared Small Target Detection

Authors: Xuelin Qian, Zixuan Wang, Chang Xu, Dingwen Zhang, Junwei Han

Abstract

Infrared small target detection (IRSTD) is a challenging task due to the extremely small size of targets, low signal-to-noise ratios, and heavy background clutter in infrared imagery. While the progress of deep learning has advanced this field, existing methods primarily adopt discriminative paradigms that directly map images to target masks, overlooking the potential of generative modeling. In this paper, we propose DiffIRSTD, a novel generative framework that reformulates infrared small target detection as a background reconstruction problem. Instead of explicitly predicting target regions, we reinterpret small targets as noise-like signals and leverage a diffusion-based model to recover a clean, target-free background from a perturbed input image. The residual between the reconstructed background and the original image is then exploited to highlight candidate target regions, providing explicit cues for subsequent detection. To address the computational cost typically associated with diffusion sampling, we introduce a one-step denoising strategy that enables efficient background reconstruction while maintaining practical inference speed. Furthermore, we present a frequency-guided perception loss to preserve high-frequency background details and focus more on the reconstruction quality of target regions. More importantly, our DiffIRSTD operates as a plug-and-play framework, compatible with existing IRSTD models without increasing the computational complexity of detectors. Extensive experiments on IRSTD-1K, NUAA-SIRST, and NUDT-SIRST datasets demonstrate the superiority of our method, achieving state-of-the-art performance across three benchmarks.

Figures & tables

Fig. 1: Illustration of our motivation. From a new perspective of denoising, (a) we treat small targets as noise-like anomalies, and (b-c) then denoise the perturbed image to obtain a target-free background image. (d) The difference between the reconstructed background and the original image can be viewed as informative clues to highlight candidate target regions, promoting the subsequent detection accuracy.

Fig. 2: Schematic of our proposed DiffIRSTD, which consists of a reconstruction network and a detection network. The reconstruction network is designed to generate a target-free background image $x_{B}$ from the input infrared image $x$ . We further incorporate a one-step denoising strategy together with a frequency-guided perceptual loss. This design not only accelerates the sampling process but also preserves the denoising quality of the reconstructed background. The detection network takes as input a dual-channel representation formed by concatenating the original infrared image and the reconstructed image. By leveraging the explicit background prior, the network can effectively perceive target regions and predict the corresponding target mask.

Fig. 3: The workflow of acquiring the pseudo-label target-free background images. Best viewed in color and zoom in.

MethodS	Backbone	IRSTD-1K			NUDT-SIRST			NUAA-SIRST
MethodS	Backbone	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
Top-Hat [ 21 ]	Traditional	10.06	75.11	1432	20.72	78.41	166.7	7.143	79.84	1012
IPI [ 22 ]	Traditional	27.92	81.37	16.18	17.76	74.49	41.23	25.67	84.63	16.67
PSTNN [ 23 ]	Traditional	24.57	71.99	35.26	14.85	66.13	44.17	30.30	72.80	48.99
MSLSTIPT [ 35 ]	Traditional	11.43	79.03	1524	8.342	47.40	888.1	10.30	82.13	1131
MTU-Net [ 36 ]	CNN + Trans.	66.11	93.27	36.80	74.85	93.97	46.95	74.78	93.54	22.36

TABLE I: Comparison with existing IRSTD approaches on IRSTD-1K, NUDT-SIRST, and NUAA-SIRST datasets. The evaluation metrics are IoU (

10^{-2}

P_{d}

(

10^{-2}

), and

F_{a}

(

10^{-6}

). The first and second best results are highlighted in bold and underlined , respectively.

\dagger

denotes methods that use diffusion models for either data augmentation, feature learning, or diffusion-based detectors.

Fig. 4: Visualization of different IRSTD methods. The green boxes indicate correctly detected targets, red boxes denote false alarms, and orange circles represent missed targets. Additionally, dashed boxes show close-up views of the targets for clearer visualization.

Detection	Reconstruction	Frequency-guided	IRSTD-1K
Network	Network	Perceptual Loss	IoU $\uparrow$	$P_{d}\uparrow$	$F_{a}\downarrow$
✓			67.13	93.93	26.64
✓	✓		69.99	92.92	25.92
✓	✓	✓	72.13	91.91	13.64

TABLE II: Ablation study on key components of DiffIRSTD.

loss variants	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
$\mathcal{L}_{mse}(\epsilon_{\theta},\epsilon)$ (Eq. 9 )	67.14	91.24	32.05
$\mathcal{L}_{mse}(x_{B},\hat{x}_{B})$ (in Eq. 18 )	69.99	92.92	25.92
$\mathcal{L}_{mse}(x_{B},\hat{x}_{B})+\mathcal{L}_{hf}$ (Eq. 18 )	72.13	91.91	13.64

TABLE III: Ablation study of reconstruction loss on IRSTD-1K.

time steps	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
30	72.16	91.58	18.95
50	72.13	91.91	13.64
70	70.69	91.58	12.63
90	71.27	93.93	14.45
120	68.94	90.90	16.13

TABLE IV: Ablation study of diffusion step

t

in Eq. 10 .

Fig. 5: Visualizations of reconstructed background images with different diffusion time steps $t$ . The red and yellow boxes denote close-ups of the foreground target and background details, respectively. Best viewed zoomed in.

In-channels	Input design	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
2	$\text{Concat}(x,\mathcal{HP}(x))$	65.47	89.22	11.06
2	$\text{Concat}(x,x_{B})$	72.13	91.91	13.64
2	$\text{Concat}(x,\|x-x_{B}\|)$	70.85	94.61	15.90
2	$\text{Concat}(x,(x-x_{B})^{2})$	71.77	90.90	13.98
3	$\text{Concat}(x,x_{B},\|x-x_{B}\|)$	69.53	93.93	21.12
3	$\text{Concat}(x,x_{B},(x-x_{B})^{2})$	69.45	93.26	21.92

TABLE V: Ablation study of dual-channel inputs on IRSTD-1K. Experiments verify the effects of background information utilization, with evaluation metrics including IoU(

10^{-2}

P_{d}

(

10^{-2}

), and

F_{a}

(

10^{-6}

). “Concat” means the concatenation operation, and “

\mathcal{HP}

” denotes the high-pass filtering operation.

$\lambda$	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
0.01	69.62	92.92	25.46
0.05	70.21	94.61	18.40
0.1	72.13	91.91	13.64
0.3	70.91	92.25	23.17
0.5	70.60	91.24	16.75
0.7	70.04	93.60	23.40

TABLE VI: Ablation study of frequency-guided perceptual loss coefficient

\lambda

on IRSTD-1K. The evaluation metrics are IoU (

10^{-2}

P_{d}

(

10^{-2}

), and

F_{a}

(

10^{-6}

$r_{c}$	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
0.1	67.74	92.92	41.50
0.3	69.87	94.27	24.50
0.5	70.26	94.94	15.88
0.7	72.13	91.91	13.64
0.9	68.40	94.27	26.64

TABLE VII: Ablation study of high-pass cutoff frequency

r_{c}

on IRSTD-1K. The evaluation metrics are IoU (

10^{-2}

P_{d}

(

10^{-2}

), and

F_{a}

(

10^{-6}

Reconstruction Network	IRSTD-1K			NUDT-SIRST			NUAA-SIRST
Reconstruction Network	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
Auto-Encoder	67.01	90.23	15.97	94.95	99.36	2.57	78.99	94.29	10.44
One-Step Diffusion	72.13	91.91	13.64	95.91	99.49	1.72	80.73	96.95	7.55

TABLE VIII: Comparisons with different reconstruction methods. The evaluation metrics are IoU (

10^{-2}

P_{d}

(

10^{-2}

), and

F_{a}

(

10^{-6}

Methods	IoU $\uparrow$	P ${}_{d}\uparrow$	F ${}_{a}\downarrow$
ACM [ 50 ]	60.98 ${}_{\text{+0.00}}$	89.89 ${}_{\text{+0.00}}$	38.26 ${}_{\text{+0.00}}$
+ DiffIRSTD ( Ours )	64.71 ${}_{\text{+3.73}}$	90.90 ${}_{\text{+1.01}}$	36.00 ${}_{\text{-2.26}}$
DNANet [ 39 ]	65.50 ${}_{\text{+0.00}}$	90.23 ${}_{\text{+0.00}}$	18.80 ${}_{\text{+0.00}}$
+ DiffIRSTD ( Ours )	69.53 ${}_{\text{+4.03}}$	91.24 ${}_{\text{+1.01}}$	14.59 ${}_{\text{-4.21}}$
SCTransNet [ 6 ]	67.28 ${}_{\text{+0.00}}$	95.28 ${}_{\text{+0.00}}$	19.49 ${}_{\text{+0.00}}$
+ DiffIRSTD ( Ours )	70.04 ${}_{\text{+2.76}}$	95.58 ${}_{\text{+0.30}}$	18.95 ${}_{\text{-0.54}}$

TABLE IX: Generalization of our DiffIRSTD with other detectors. The red number means the performance gain on IRSTD-1K.

Infrared Small Target Detection

Infrared small target detection finds faint, often sub-pixel objects against cluttered thermal backgrounds while limiting false alarms.

See all 43 papers

Denoising-Enhanced Coarse-to-Fine Infrared Small Target Detection with Attention Prior-Guided Knowledge Distillation
Jun 20, 2026Houzhang Fang, Ruixuan Huang, Qiuhuan Chen +3Attention-Guided Distillation Infrared Small Target Detection
Xidian University, Xi’an, China · Huazhong University of Science and Technology, Wuhan, China
FSGNet: A Frequency-Aware and Semantic Guidance Network for Infrared Small Target Detection
Mar 26, 2026Yingmei Zhang, Wangtao Bao, Yong Yang +3Infrared Small Target Detection Object Detection
School of Software and Internet of Things Engineering, Jiangxi University of Finance and Economics, Nanchang 330032, China · the postdoctoral researcher, CGN Begood Technology co., Ltd, Nanchang 330032, China · School of Computer Science and Technology, Tiangong University, Tianjin 300387, China +1
DGNet: Dual-knowledge Guided Network for Infrared Small Target Detection
Sep 1, 2026Chenglong Yu, Mingzhu Xu, Jing Wang +3Infrared Small Target Detection Object Detection
Shandong University · Jinan, China · Harbin Institute of Technology, Shenzhen +1

Jun 20, 2026J/K move · Enter open · S save

Denoising-Enhanced Coarse-to-Fine Infrared Small Target Detection with Attention Prior-Guided Knowledge Distillation

Houzhang Fang, Ruixuan Huang, Qiuhuan Chen, Xiaolin Wang+2

Xidian University, Xi’an, China · Huazhong University of Science and Technology, Wuhan, China

Attention-Guided Distillation Infrared Small Target Detection Object Detection Real-Time Object Detection Knowledge Distillation

Open paper AskPDF arXiv

Infrared small target detection (IRSTD) in high-resolution images is crucial for unmanned aerial vehicle (UAV) surveillance and UAV-based ground monitoring. However, small target size, weak features, and interference from complex dynamic backgrounds make IRSTD challenging. Existing methods incur redundant computation in non-target background regions and insufficiently exploit target context, limiting detection performance. To address these issues, we propose ECFNet, an efficient coarse-to-fine IRSTD framework with attention prior-guided knowledge distillation. In the coarse stage, we design a region binary classification network (RBCN) on grid-based multi-scale feature maps to efficiently identify target-containing context region proposals. A new denoising-assisted training strategy incorporates noisy ground-truth (GT) masks into RBCN feature maps and trains the network to reconstruct the original GT masks. This auxiliary task encourages explicit learning of target-background context to better distinguish target proposals from background regions. In the fine stage, we customize a lightweight target detector to the coarse-stage region proposals to balance accuracy and efficiency. Furthermore, we introduce a knowledge distillation strategy guided by a teacher-student cross-attention prior. This strategy directs the student to focus on critical target regions, enhancing discriminative feature representations for infrared small targets. Extensive experiments on three real infrared datasets demonstrate that ECFNet outperforms existing single-stage and two-stage approaches while maintaining high real-time processing efficiency. Code: https://github.com/IVPLabs/ECFNet.

Abstract

Infrared Small Target Detection

Explore similar work

Denoising-Enhanced Coarse-to-Fine Infrared Small Target Detection with Attention Prior-Guided Knowledge Distillation

FSGNet: A Frequency-Aware and Semantic Guidance Network for Infrared Small Target Detection

DGNet: Dual-knowledge Guided Network for Infrared Small Target Detection